[spullara]

I've always wondered why completion i/o systems require you to hand them a buffer rather than a buffer pool. Handing them a buffer means that you are keeping a huge amount of memory allocated when you have a huge number of connections.

[h33p]

For TCP, you'd generally want to have a buffer pool and issue one repeating request that takes buffers out from that pool. This is what io_uring allows you to do.

The difficulty comes from APIs ability to expose that buffer pool nicely. I am on the lookout to model this better, but no clear solution that does not make assumptions about the backend at play.

For now, however, with careful design you can skip the allocated buffers and supply a stack reference. So long as you have assurances the stack wont move (Rust's pinning guarantee), you should be able to pass that reference to the io system without problems. Not saying this results in zero copies being made - the backend usually has strict alignment requirements and alike, but hey, we at least get rid of one allocation!

[infamouscow]

To add, if your app can make use of io_uring's polling, you can further avoid a copy and have a true zero-copy system.

[wtallis]

io_uring appears to have had this feature since kernel 5.7 in 2020.